Wafer raw material usage cost prediction management method and system based on work order feature learning

Abstract

The application relates to the technical field of wafer manufacturing resource planning and prediction management, and discloses a wafer raw material consumption cost prediction management method and system based on work order feature learning, which comprises the following steps: calling a target work order and generating a standard work order feature vector; opening a historical work order time sequence window and loading previous historical work order features, generating a time sequence state inertia operator through discrete difference and time attenuation weighting; mapping the feature vector through a basic correlation coefficient matrix and introducing the operator to recombine and generate an adaptive contribution weight matrix; calculating material benchmark demand and adding a safety correction coefficient to output a total raw material consumption, and sending a message when the absolute deviation value of the total amount and the inventory digital component exceeds the threshold value of 5% for three times in succession, the wafer raw material consumption cost prediction management method and system based on work order feature learning can offset the disturbance of production change through the time sequence state inertia operator, guide the prediction to dynamic time sequence cascade prediction transition, and improve the system stability.

Description

Technical Field

[0001] This invention relates to a method and system for predicting and managing the consumption costs of wafer raw materials based on work order feature learning, belonging to the field of wafer manufacturing resource planning and forecasting management technology. Background Technology

[0002] Currently, establishing raw material turnover and cost control plans based on historical production records is a common management approach. This approach relies on data processing centers to retrieve multi-feature historical scheduling work orders from the manufacturing system, analyze the material allocation patterns of different wafer specifications, generate a raw material usage forecast flow within the standard production cycle, and issue automated procurement instructions and balance warehouse inventory accordingly. The wafer raw material allocation share is composed of multi-dimensional work order feature variables, which exhibit a highly dynamic and strongly nonlinear distribution in the multi-dimensional feature space. The transient offset trajectory of raw material consumption rate corresponds to the complex coupling of specific production machine status, wafer physical specifications, and basic process parameters within the discrete manufacturing cycle. However, when facing continuous scheduling scenarios involving high-frequency customized product changeovers, existing management methods treat each production work order as an independent static data entity in the time series, failing to detect the inertia of operating conditions such as residual thermodynamic gradients and fluctuations in chemical gas saturation that are common in physical machines during continuous changeovers. This causes the prediction weights calculated by the system within the changeover transition window to drift nonlinearly, resulting in prediction data deviating from the actual material consumption trajectory and generating material supply and demand data deviations during peak scheduling periods.

[0003] Faced with the continuous scheduling of high-frequency customized product changeovers, existing management methods treat production work orders as independent, static data entities in a time series. They fail to recognize the inertia of operating conditions such as residual thermodynamic gradients and fluctuations in chemical gas saturation during continuous physical machine changeovers. This causes nonlinear drift in the prediction weights generated within the system's changeover transition window, resulting in predicted data deviating from the actual material consumption trajectory. Besides hardware-level physical operating condition fluctuations, existing software control and prediction methods also have shortcomings at the algorithmic logic level. For example, Chinese invention patent application CN120406056A discloses a method, device, storage medium, and equipment for predicting photoresist consumption. It improves prediction accuracy by fitting the nonlinear relationship between wafer input and consumption using a polynomial regression model, but in actual production processes… In this approach, the solution relies on macroscopic statistics of the total number of wafers put into production within a preset time period, failing to penetrate and characterize the time-series coupling characteristics of adjacent work orders in continuous scheduling. Due to the lack of digital capture of the inertial evolution of machine status, the regression logic exhibits significant perception lag when faced with transient shifts caused by frequent production changes, making it difficult to maintain real-time accuracy of material control in dynamic production environments. If the above drift is forcibly fitted by increasing the iteration frequency of the feature vector regression operation, it will trigger repeated inversion of the high-dimensional space matrix, causing processor memory bus throughput overload and resulting in system bus response delay exceeding 350ms. If the initial disturbance data of track change is simply cleared, the calculation model will face the risk of algorithm divergence due to the lack of sufficient time-series flow when facing data sparse conditions such as the number of valid historical work order records within the historical sequence window being less than 5.

[0004] Therefore, the technical problem to be solved by this invention is how to utilize the existing digital work order records in the system to reshape the time series constraint mechanism within the digital system to offset the residual interference of continuous production changeover in the physical workshop, and maintain the convergence stability and low latency response of the management system under non-ideal environments such as data sparsity or excessive computing power load. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A method for predicting and managing the cost of wafer raw material usage based on work order feature learning, comprising the following steps: Step S101: Retrieve the target work order from the data warehouse, extract the physical wafer area, the number of designed exposure layers, and the equipment planned capacity components from the target work order, normalize them into values ​​between 0 and 1, and generate a standard work order feature vector. Step S102: Establish a historical work order time sequence window with a depth of 5 to 12 in memory, load the standard work order feature vector of the previous historical work order of the target work order to construct a time sequence matrix, and generate a time sequence state inertia operator that characterizes the trend of process disturbance by performing discrete difference and time decay weighted calculation. Step S103: Apply the basic correlation coefficient matrix to perform multiplication mapping on the standard work order feature vector, and introduce the time-series state inertia operator as a dynamic compensation quantity to generate an adaptive contribution weight matrix. The components in the adaptive contribution weight matrix are the sum of the basic correlation coefficient and the product term in the corresponding dimension, and the product term is the value of the product of the attenuation coefficient and the component of the time-series state inertia operator in the corresponding dimension plus 1. Step S104: Calculate the inner product of the adaptive contribution weight matrix and the standard work order feature vector to obtain the material baseline requirement, accumulate a safety correction coefficient between 1.04 and 1.12 to output the total expected raw material consumption, and send a data message containing the order quantity and scheduling instructions to the external procurement management response module when the absolute deviation between the total expected raw material consumption and the digital inventory component exceeds the 5% threshold three times consecutively.

[0006] Preferably, step S102 includes the following sub-steps: Step S1021, retrieve the historical work order sample set from the storage unit, and count the number of historical data points in the historical work order sample set that match the feature vector of the standard work order; Step S1022, when the number of historical data points is less than 5, stop calculating the timing state inertia operator, and retrieve the process similarity matrix from the storage unit; Step S1023, extract the average value of the target historical section with the highest similarity to the target work order from the process similarity matrix as the replacement benchmark, and replace the timing state inertia operator with the replacement benchmark, while forcibly locking the attenuation coefficient to a fixed constant of 0.03.

[0007] Preferably, step S101 includes the following sub-steps: step S1011, extracting the physical wafer area, the number of designed exposure layers, and the equipment planned capacity components from the target work order, and combining them to form a work order parameter group; step S1012, performing range normalization mapping on the work order parameter group, and converting it to generate a standard work order feature vector.

[0008] Preferably, step S103 includes the following sub-steps: step S1031, reading the bus response delay of the central processing unit and obtaining the queue of requests to be processed; step S1032, when the bus response delay is greater than 350ms, reading the material price feature value corresponding to each work order in the queue of requests to be processed; step S1033, for work order requests whose material price feature value is lower than a set threshold, reducing the dimensionality of their feature vectors and retaining only 3 feature dimensions so that the bus response delay is no greater than 400ms.

[0009] Preferably, in step S103, the generation of the adaptive contribution weight matrix includes the following sub-steps: step S1034, calculate the contribution weight of each feature item to the amount of wafer raw materials based on the variance distribution of each feature item in the historical work order sample set; step S1035, perform matrix mapping using the contribution weights to generate the adaptive contribution weight matrix.

[0010] Preferably, step S104 includes the following sub-steps: step S1041, when the absolute deviation value exceeds the 5% threshold three times consecutively, the adaptive contribution weight matrix is ​​corrected by introducing a residual correction term; step S1042, the current inventory digital component is extracted and the inventory shortage is calculated, a data message containing the purchase order quantity and emergency dispatch instructions is generated and sent to the external procurement management response module.

[0011] Preferably, in step S1022, retrieving the stored process similarity matrix includes the following sub-steps: Step S10221, retrieving parameters of historical wafer work orders in the storage unit, calculating the similarity between each historical wafer work order, and recombining to generate a process similarity matrix; Step S10222, calculating the cosine similarity of the angle between the target work order and each historical wafer work order in the process similarity matrix, and determining the top 5 historical wafer work orders with the highest similarity.

[0012] Preferably, in step S104, driving the decision state machine transition includes the following sub-steps: Step S1043, establishing and initializing the decision state machine, setting the initial state of the decision state machine to the normal prediction state; Step S1044, when a trigger signal is received that the absolute deviation value exceeds the 5% threshold three times consecutively, controlling the decision state machine to transition from the normal prediction state to the emergency scheduling state, triggering the generation and sending of data packets.

[0013] Preferably, in step S102, loading the standard work order feature vector of the preceding historical work orders of the target work order includes the following sub-steps: step S1024, obtaining the deviation value between the actual total raw material consumption of the completed work order and the initial predicted consumption; step S1025, storing the actual total raw material consumption and the deviation value in the storage unit to update the historical work order sample set.

[0014] A wafer raw material usage cost prediction and management system based on work order feature learning, which is used to implement a wafer raw material usage cost prediction and management method based on work order feature learning, including: The feature generation module is used to retrieve target work orders from the data warehouse, extract the physical wafer area, the number of designed exposure layers, and the equipment planned capacity components from the target work orders, normalize them into values ​​between 0 and 1, and generate standard work order feature vectors. The timing analysis module is used to create a historical work order timing window with a depth of 5 to 12 in memory, load the standard work order feature vector of the previous historical work orders of the target work order to construct a timing matrix, and generate a timing state inertial operator that characterizes the trend of process disturbance through discrete difference and time decay weighted calculation. The weight adaptive control module, connected to the feature generation module and the time series analysis module, is used to apply the basic correlation coefficient matrix to perform multiplication mapping on the standard work order feature vector, and introduce the time series state inertia operator as a dynamic compensation quantity to generate an adaptive contribution weight matrix. The decision scheduling management module, connected to the feature generation module and the weight adaptive control module, is used to calculate the inner product of the adaptive contribution weight matrix and the standard work order feature vector to obtain the material baseline demand, accumulate a safety correction coefficient between 1.04 and 1.12 to output the total expected raw material consumption, and send a data message containing the order quantity and scheduling instructions to the external procurement management response module when the absolute deviation between the total expected raw material consumption and the digital inventory component exceeds the 5% threshold for three consecutive times.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In the cost prediction and management of wafer raw material usage, the data processing unit opens a forward sliding time sequence window, retrieves the completed historical feature vectors of the preceding sequence to construct a time sequence state space matrix, and obtains a time sequence state inertia operator by combining discrete difference and time distance attenuation weighted operation. In the mapping loop, this operator is used as a dynamic compensation factor to reorganize and generate an adaptive contribution weight matrix. This mechanism changes the conventional static parameter individual regression mode, and transforms the process switching transition state between successive production batches in the time series into a computable data structure, so that the feature weights of each dimension generate adaptive displacement with the scheduling flow, offsetting the nonlinear fluctuation of raw material consumption rate caused by continuous production change of physical equipment, and ensuring the stability of weight calculation of the prediction system under the condition of high frequency production change of multiple varieties.

[0016] 2. The system is configured with a data reshaping state machine, which incorporates a fallback hedging logic for non-ideal operating conditions during time-series calculations. When the number of valid completed work orders within the pre-sliding time-series window is less than 5, the state machine immediately cuts off the dynamic hedging loop of the weight matrix, automatically retrieves the preset global process similarity topology map from the storage unit, selects the average of typical historical sections with the highest affinity to the current target work order as the replacement benchmark, and simultaneously forcibly locks the discrete inertia decay coefficient to a fixed boundary constant of 0.03. This mechanism effectively solves the data sparsity bottleneck caused by the launch of new processes or the truncation of historical databases, avoids weight divergence in the calculation model when there is a lack of sufficient historical time-series flow, and maintains the algorithm convergence stability and operational boundary safety of the management prediction system under extreme edge conditions.

[0017] 3. The system incorporates a dynamic memory queue arbitration module at the computation output end to handle the computing power-constrained conditions caused by the concentrated influx of multiple concurrent scheduling prediction requests during peak production change periods. The arbitration module reads the transient thermal power drift of the processor core and the bus response latency in real time. When the latency exceeds 350ms, a heuristic screening process is initiated. The material cost index corresponding to each prediction request in the queue is extracted as a priority weighting sorting factor. For low-priority requests with lower material unit prices, the dimension of their feature vectors is forcibly reduced from high-dimensional space discretization to core 3D features. This scheduling logic, starting from the coordinated change of memory throughput and high-dimensional computing, effectively reduces the memory throughput pressure of the central processing unit when multiple tasks are parallel and change frequently, and forcibly truncates and controls the system bus latency within the 400ms safety line. Attached Figure Description

[0018] Figure 1 This is a flowchart of the predictive management method including time-series state inertia compensation according to the present invention; Figure 2 This is a module communication topology diagram of the wafer raw material cost prediction and management system of the present invention.

[0019] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0021] A method for predicting and managing the consumption costs of wafer raw materials based on work order feature learning includes the following steps: Step S101: Retrieve the target work order from the data warehouse, extract the physical wafer area, the number of designed exposure layers, and the equipment planned capacity components from the target work order, normalize them into values ​​between 0 and 1, and generate a standard work order feature vector. Step S102: Establish a historical work order time sequence window with a depth of 5 to 12 in memory, load the standard work order feature vector of the previous historical work order of the target work order to construct a time sequence matrix, and generate a time sequence state inertia operator that characterizes the trend of process disturbance by performing discrete difference and time decay weighted calculation. Step S103: Apply the basic correlation coefficient matrix to perform multiplication mapping on the standard work order feature vector, and introduce the time-series state inertia operator as a dynamic compensation quantity to generate an adaptive contribution weight matrix. The components in the adaptive contribution weight matrix are the sum of the basic correlation coefficient and the product term in the corresponding dimension, and the product term is the value of the product of the attenuation coefficient and the component of the time-series state inertia operator in the corresponding dimension plus 1. Step S104: Calculate the inner product of the adaptive contribution weight matrix and the standard work order feature vector to obtain the material baseline requirement, accumulate a safety correction coefficient between 1.04 and 1.12 to output the total expected raw material consumption, and send a data message containing the order quantity and scheduling instructions to the external procurement management response module when the absolute deviation between the total expected raw material consumption and the digital inventory component exceeds the 5% threshold three times consecutively.

[0022] Preferably, step S102 includes the following sub-steps: Step S1021, retrieve the historical work order sample set from the storage unit, and count the number of historical data points in the historical work order sample set that match the feature vector of the standard work order; Step S1022, when the number of historical data points is less than 5, stop calculating the timing state inertia operator, and retrieve the process similarity matrix from the storage unit; Step S1023, extract the average value of the target historical section with the highest similarity to the target work order from the process similarity matrix as the replacement benchmark, and replace the timing state inertia operator with the replacement benchmark, while forcibly locking the attenuation coefficient to a fixed constant of 0.03.

[0023] Preferably, step S101 includes the following sub-steps: step S1011, extracting the physical wafer area, the number of designed exposure layers, and the equipment planned capacity components from the target work order, and combining them to form a work order parameter group; step S1012, performing range normalization mapping on the work order parameter group, and converting it to generate a standard work order feature vector.

[0024] Preferably, step S103 includes the following sub-steps: step S1031, reading the bus response delay of the central processing unit and obtaining the queue of requests to be processed; step S1032, when the bus response delay is greater than 350ms, reading the material price feature value corresponding to each work order in the queue of requests to be processed; step S1033, for work order requests whose material price feature value is lower than a set threshold, reducing the dimensionality of their feature vectors and retaining only 3 feature dimensions so that the bus response delay is no greater than 400ms.

[0025] Preferably, in step S103, the generation of the adaptive contribution weight matrix includes the following sub-steps: step S1034, calculate the contribution weight of each feature item to the amount of wafer raw materials based on the variance distribution of each feature item in the historical work order sample set; step S1035, perform matrix mapping using the contribution weights to generate the adaptive contribution weight matrix.

[0026] Preferably, step S104 includes the following sub-steps: step S1041, when the absolute deviation value exceeds the 5% threshold three times consecutively, the adaptive contribution weight matrix is ​​corrected by introducing a residual correction term; step S1042, the current inventory digital component is extracted and the inventory shortage is calculated, a data message containing the purchase order quantity and emergency dispatch instructions is generated and sent to the external procurement management response module.

[0027] Preferably, in step S1022, retrieving the stored process similarity matrix includes the following sub-steps: Step S10221, retrieving parameters of historical wafer work orders in the storage unit, calculating the similarity between each historical wafer work order, and recombining to generate a process similarity matrix; Step S10222, calculating the cosine similarity of the angle between the target work order and each historical wafer work order in the process similarity matrix, and determining the top 5 historical wafer work orders with the highest similarity.

[0028] Preferably, in step S104, driving the decision state machine transition includes the following sub-steps: Step S1043, establishing and initializing the decision state machine, setting the initial state of the decision state machine to the normal prediction state; Step S1044, when a trigger signal is received that the absolute deviation value exceeds the 5% threshold three times consecutively, controlling the decision state machine to transition from the normal prediction state to the emergency scheduling state, triggering the generation and sending of data packets.

[0029] Preferably, in step S102, loading the standard work order feature vector of the preceding historical work orders of the target work order includes the following sub-steps: step S1024, obtaining the deviation value between the actual total raw material consumption of the completed work order and the initial predicted consumption; step S1025, storing the actual total raw material consumption and the deviation value in the storage unit to update the historical work order sample set.

[0030] A wafer raw material usage cost prediction and management system based on work order feature learning includes: The feature generation module is used to retrieve target work orders from the data warehouse, extract the physical wafer area, the number of designed exposure layers, and the equipment planned capacity components from the target work orders, normalize them into values ​​between 0 and 1, and generate standard work order feature vectors. The timing analysis module is used to create a historical work order timing window with a depth of 5 to 12 in memory, load the standard work order feature vector of the previous historical work orders of the target work order to construct a timing matrix, and generate a timing state inertial operator that characterizes the trend of process disturbance through discrete difference and time decay weighted calculation. The weight adaptive control module, connected to the feature generation module and the time series analysis module, is used to apply the basic correlation coefficient matrix to perform multiplication mapping on the standard work order feature vector, and introduce the time series state inertia operator as a dynamic compensation quantity to generate an adaptive contribution weight matrix. The decision scheduling management module, connected to the feature generation module and the weight adaptive control module, is used to calculate the inner product of the adaptive contribution weight matrix and the standard work order feature vector to obtain the material baseline demand, accumulate a safety correction coefficient between 1.04 and 1.12 to output the total expected raw material consumption, and send a data message containing the order quantity and scheduling instructions to the external procurement management response module when the absolute deviation between the total expected raw material consumption and the digital inventory component exceeds the 5% threshold for three consecutive times.

[0031] Example 1: In an integrated circuit manufacturing resource planning environment that includes production changeover scheduling for multiple types of customized wafers, when the system faces a series of high-frequency, continuous customized production orders with high exposure layers, causing fluctuations in the saturation of the gas remaining in the reaction chambers of the lithography and etching equipment, subsequent mass production or R&D wafer orders will experience sudden nonlinear shifts in the raw material consumption rate. Simultaneously, the concentrated influx of multiple concurrent scheduling instructions from multiple workshops within a short period leads to CPU memory throughput overload and bus response latency exceeding 350ms. Furthermore, during the initial stages of production changeover or the rollout of a new process, a leading sliding timing window is often present. In situations where the number of completed work orders is less than 5, data sparsity leads to traditional static forecasting management systems failing to digitally represent the time-span migration inertia between sequential work orders. This results in material supply and demand mismatches, algorithm divergence, and large-scale inventory backlogs. In contrast, a wafer raw material usage cost forecasting management system based on work order feature learning, upon receiving a target work order to be scheduled, uses a feature generation module to retrieve the target work order from the data warehouse. It extracts the physical wafer area, the number of designed exposure layers, and the planned equipment capacity components, and normalizes them through range normalization mapping to convert them into values ​​between 0 and 1, generating a standard work order feature vector. The synchronous timing analysis module establishes historical work order timing windows in memory with depths of 5 to 12, currently set to 8. If the number of completed valid historical work order records within the historical work order timing window is less than 5, the data reshaping state machine immediately cuts off the dynamic hedging loop of the weight matrix and retrieves the preset global process similarity topology map from the storage unit. It calculates the cosine similarity of the angle between the target work order and each historical wafer work order in the process similarity topology map to determine the top 5 historical wafer work orders with the highest similarity. It extracts the average value of the corresponding typical historical cross-section as a replacement benchmark to replace the timing state inertia operator and the discrete inertia decay coefficient. The boundary constant is locked at 0.03. Under normal operating conditions where the number of valid historical work order records is greater than or equal to 5, the standard work order feature vector of the previous historical work order is directly loaded to construct the temporal state space matrix. Combined with discrete difference and time distance attenuation weighted calculation, a temporal state inertia operator characterizing the residual fluctuations of the process environment and the trend of process disturbances is generated. Based on the theory of capturing transient change rate and exponential decay model of first-order difference to characterize the time-dependent physical residual effect, the weighted calculation of discrete difference and time distance decay adopts the following: Among them, the temporal state inertial operator in the th... Numerical components in each feature dimension The time-distance decay constant takes a value between -1 and +1. The value ranges from 0.1 to 0.5, representing the historical work order time sequence window depth. The value is between 5 and 12. From 1 to Discrete time step index, The base of the natural logarithm, and Each is the first in the timing window The and the first The historical work order is in the [number]th [number]. Normalized feature components with values ​​between 0 and 1 on each feature dimension, where all parameters and components are dimensionless pure numbers. This is the feature vector of a standard work order. For the normalized feature components of each dimension, The total number of feature dimensions. The discrete inertial attenuation coefficient is a dimensionless pure number. A time-series state inertial operator to characterize the trend of process disturbances. These are the numerical components of the temporal state inertia operator in each feature dimension.

[0032] In the feature weight reorganization calculation loop, to prevent computational crashes caused by multiple concurrent high-frequency production change scheduling requests, the system's back-end dynamic memory queue arbitration module reads the transient thermal power drift of the central processing unit and the bus response latency in real time. When the bus response latency is determined to be greater than 350ms, the module reads the material price feature value corresponding to each work order in the pending request queue, and for low-priority work order requests with material price feature values ​​lower than a set threshold, forces the dimension of their feature vectors to be changed from... This invention reduces memory throughput and keeps system bus latency within a safe 400ms limit by discretizing and downgrading the three feature dimensions. When processing high-dimensional features, the CPU needs to perform a large number of matrix inversions and multiply-accumulate operations at high frequency, which leads to a sharp increase in data transfer throughput on the memory bus and triggers bus contention. This invention addresses this by forcibly reducing the feature vector dimension of low-priority work orders from... By discretizing and reducing the dimensions to the core three feature dimensions, unnecessary addressing channels for other high-dimensional matrix operations are closed. This reduces the number of multiply-accumulate instruction cycles and memory addressing block size for a single inner product operation by more than 60%, fundamentally eliminating bus queuing deadlocks caused by data backlog. This allows the physical system bus response latency to quickly decrease and be forcibly truncated and controlled. Within the security defense line, the weighted adaptive control module applies the pre-stored basic correlation coefficient matrix in the memory. For the feature vector of standard work orders Implement multiplication mapping and simultaneously introduce the temporal state inertial operator. As a dynamic compensation quantity woven into the loop, an adaptive contribution weight matrix incorporating temporal state transition constraints is calculated and generated. The formula used Determine the adaptive contribution weight matrix The components in, among which, For the current target work order in the 1st Adaptive contribution weights on each feature dimension For the pre-stored number The basic correlation coefficients of the features in each dimension This is the discrete inertial attenuation coefficient, set to 0.05 under normal conditions. For the time-series state inertial operator in the th Numerical components in each feature dimension From 1 to The feature dimension corresponds to the index. Based on the basic correlation coefficient matrix, The adaptive contribution weight matrix.

[0033] The decision scheduling management module calculates the adaptive contribution weight matrix. With standard work order feature vector The inner product is used to obtain the baseline material requirement component, and a fixed safety correction factor of 1.08, dynamically linked to historical consumption residuals, is simultaneously added to calculate and output the total estimated material consumption for wafer raw materials. In practical applications, the quantification range of the safety correction coefficient is limited to between 1.04 and 1.12. The boundary selection is based on the following: if the safety correction coefficient is below 1.04, the system lacks sufficient safety buffer redundancy to handle transient excess material consumption caused by sudden production changes, easily leading to production line shutdowns due to material shortages; if it is above 1.12, it will result in over-purchasing, leading to redundant material stockpiling at the warehousing end and increasing material stagnation costs. The specific rule for dynamically linking it to historical consumption residuals is as follows: the system's calculation and processing unit calculates the mean square error between the actual usage and the initial predicted usage of several completed work orders within the historical work order time window in real time. When this mean square error increases, the safety correction coefficient is finely adjusted upwards within the linear range to improve the system's ability to defend against material shortages; conversely, it is finely adjusted downwards to approximate the theoretical material consumption line. Then, the corresponding inventory digitization component in the material storage register is read and the absolute deviation between the two is calculated. Simultaneously, the decision state machine is initialized, and the initial state is set to the normal prediction state. When the total expected raw material consumption... When the absolute deviation from the digitized inventory value exceeds the preset safety stock fluctuation threshold of 5% three consecutive times, the control decision state machine transitions from the normal prediction state to the emergency dispatch state. It automatically encapsulates and generates a digital data message containing the purchase order quantity and emergency dispatch instructions, and sends it directly to the external procurement management response module to implement supply chain resource allocation management. Based on the material balance conservation and inventory feedback closed-loop control theory, after receiving the data message, the external procurement management response module processing unit analyzes the expected total raw material consumption and the digitized inventory value, and determines the purchase order quantity using the following formula: Among them, the purchase order quantity Greater than 0, estimated total consumption of raw materials Greater than 0, the digitized inventory component read from the material storage register. Greater than or equal to 0, preset safety reserve component For variables greater than 0, all variables use the specific material's nominal unit; the external procurement management response module calculates the purchase order quantity. The system sends an electronic supply chain procurement offer instruction containing material specification codes, purchase quantities, and expected delivery times to an external supplier server. This instruction drives the electrically controlled proportional feeder valve in the production line's batching bin to adjust its opening, aligning the valve opening with the purchase order quantity. The correlation is monotonically positive. Changes in physical adjustment scheduling cause inventory fluctuations. At the same time, the system applies residual correction terms to the adaptive contribution weight matrix. Dynamic updates are implemented to complete weight compensation, enabling in-situ offsetting of nonlinear disturbances caused by the drift of upstream equipment status, reducing material inventory turnover cycle, and eliminating material shortages, shutdowns, and warehouse backlogs. Specifically, the residual correction term is characterized by the relative residual between the actual total raw material consumption and the expected total raw material consumption of the current completed work order. The correction calculation process is as follows: calculate the difference between the actual total raw material consumption and the expected total raw material consumption and divide it by the actual total raw material consumption to obtain the relative deviation rate. Use this relative deviation rate as the feedback adjustment gain and multiply it by each weight component in the original adaptive contribution weight matrix to obtain the updated adaptive weight components. This enables in-situ fine-tuning and dynamic updating of the adaptive contribution weight matrix W, ensuring the adaptive convergence of the predicted weights for subsequent work orders.

[0034] Example 2: A distributed simulation platform for wafer manufacturing resource planning based on high-frequency trajectory scheduling is configured with multi-core computing nodes with a main frequency of 3.2GHz and 64GB of memory. Its simulation clock precision is set to 0.1ms, used to provide the feature generation module with wafer area ranging from 100mm². 2 Up to 300mm 2The design incorporates 2000 irregularly arranged structured historical work order data streams from wafer manufacturing, ranging from 10 to 60 layers. During simulation, irregular data noise with an amplitude of 3.5% of the original signal component is actively injected into the input original historical work order feature stream via a data stream overlay component. Simultaneously, a random response delay disturbance with an average value of 45ms and a peak value of 380ms is loaded into the internal system bus via a bus delay generator to construct a highly realistic test environment. The depth setting of the historical work order timing window corresponding to the timing analysis module is determined by the complete timing feature capture. The physical constraints between performance and matrix operation overhead dictate that when the window depth is set too small (below the lower boundary of 5 cycles), the computational loop cannot fully extract the gas saturation fluctuation trend remaining in the reaction chamber from previous high-exposure layer work orders, leading to underfitting. Conversely, when the window depth is set too large (above the upper boundary of 12 cycles), the matrix multiplication overhead when reconstructing the adaptive contribution weight matrix increases exponentially, causing the processor's computing power to exceed its limit and frequently triggering low-priority feature pruning, consuming system hardware addressing resources. Based on these technical trade-offs... In this experiment, the nominal depth of the historical work order time sequence window was selected as 8. To evaluate the anti-interference stability of the method under a multi-task parallel high-frequency trajectory change environment and to verify the rationality of the numerical boundary, a control system was constructed, including the present invention sample group, the control group, the partially missing control group 1, the partially missing control group 2, and the out-of-range control group 1 and the out-of-range control group 2. The present invention sample group was configured with a historical work order time sequence window with a nominal depth of 8 and enabled a complete adaptive weight compensation loop. The control group adopted a static material prediction control algorithm based on the historical cross-sectional average value. The partially missing control group 1 shut down the dynamic hedging path of the time sequence state inertia operator when reorganizing and generating the adaptive contribution weight matrix. The partially missing control group 2 stopped the data reshaping state machine when the number of valid historical work order records completed within the sliding time sequence window was less than 5. The out-of-range control group 1 and the out-of-range control group 2 respectively forced the depth of the historical work order time sequence window to be 4 and 13, respectively, which deviated from the statutory quantization window. At the same time, a problem intensity gradient control system covering three levels of low-complexity, medium-complexity, and high-complexity processes was introduced to observe the total expected consumption of raw materials output by the prediction system. The deviation and fluctuation pattern of the digital inventory component in the material storage register under different scheduling saturation impacts.

[0035] The test data from 1000 discrete manufacturing cycles continuously running on a distributed simulation platform shows that, under low-complexity process gradients, the absolute deviation of cost prediction in the control group remained at 8.34%, while the absolute deviation of cost prediction in the present invention's sample group remained at 1.12%. As the problem intensity gradient extends to high-complexity process conditions involving continuous multi-product changes, the absolute deviation of prediction in the control group significantly increases to 22.67% due to its inability to track nonlinear material consumption rate mutations caused by cavity residue. Furthermore, concurrent scheduling requests from multiple workshops lead to a system bus response latency of 384.6ms, resulting in a processing queue deadlock. In contrast, under the same high-complexity process conditions, the present invention's sample group, through its dynamic memory queue arbitration module, performs dimensionality reduction and pruning for low-priority work orders with material price characteristic values ​​below a set threshold, keeping the system bus response latency within a safe range of 368.2ms. Its total estimated raw material consumption... The absolute prediction deviation remained stable at 2.45%. Compared with the data from the partially missing control group, the absolute prediction deviation increased to 14.18% under high-complexity process conditions, confirming the effectiveness of the time-series state inertia operator. In the formula The necessity of applying dynamic displacement hedging to the feature weights, whereby... For adaptive contribution weight matrix, For the current target work order in the 1st Adaptive contribution weights on each feature dimension Based on the basic correlation coefficient matrix, For the pre-stored number The basic correlation coefficients of the features in each dimension The discrete inertial attenuation coefficient is a dimensionless pure number and is set to 0.05. For the time-series state inertial operator in the th Numerical components in each feature dimension From 1 to The feature dimensions correspond to subscripts. Furthermore, in the non-ideal data sparsity condition where the number of completed records is less than 5 due to artificially truncated historical work order streams, the second partial missing control group experienced a divergence in its computational model due to the breakage of the temporal state matrix, resulting in an uncontrolled expansion of its absolute prediction deviation to 31.56%. In contrast, the sample group of this invention uses a data reshaping state machine to access a preset global process similarity topology map, extracts the average of typical historical cross-sections of the top 5 wafer work orders based on the cosine similarity of the included angle, and uses the discrete inertia decay coefficient... By forcibly locking the boundary constant to 0.03, the absolute deviation of the prediction value was kept at 3.88%, verifying the operational reliability under the environment of missing local features. The nonlinear law analysis of the extreme boundary of the time-series numerical window revealed that when the depth of the historical work order time-series window was forcibly assigned to 4 in the out-of-range control group, the time-series span at the component level could not fully cover the long cycle of process migration caused by the high exposure layer production change of the physical machine, resulting in the generation of the time-series state inertia operator. The lack of ability to characterize the cascading evolution of residual material saturation resulted in a sharp increase in the absolute deviation of material prediction to 16.74% under high-complexity processes. However, when the window depth of the out-of-range control group was set to 13, the absolute deviation of prediction remained at 2.41%. Compared with the sample data of this invention with a depth of 8, no statistically significant improvement in prediction accuracy was observed. This indicates that the saturation point of the system's time-series feature capture information is located at a window depth of 12. Furthermore, due to the participation of large matrices in multi-cycle inner products and multiplication recombination, the high-dimensional feature flow severely consumes the physical memory of the controller node, causing the system bus response delay to rise sharply to 412.3ms, thus breaking through the 400ms system safety limit. This nonlinear performance inflection point directly serves as a quantitative factual argument for determining that the time-series window depth in the technical solution of this invention is within the range of 5 to 12 as an interval limit, proving that the legally quantified range belongs to the global working range that takes into account both the response time of resource planning and the accuracy of material control.

[0036] Based on the multidimensional quantitative technical facts produced by the above-mentioned comparative experiments, the calculation and processing unit of the present invention offsets the interference of the sequence transition state under the continuous track change scheduling condition in situ. It converts the process switching transition state between successive production batches in the time series into a computable time-series state inertial operator. This allows the normalized feature weights of each dimension to undergo adaptive dynamic shifts as scheduling instructions flow, predicting the total estimated raw material consumption output. Automatic high-precision alignment with digital inventory components reduces administrative delays and redundant material backlogs in the wafer manufacturing process. When the absolute deviation between the two exceeds the 5% inventory safety limit three times consecutively, the control decision state machine automatically transitions from the normal prediction state to the emergency scheduling state, automatically sending digital data messages containing deterministic material procurement and allocation instructions to the external procurement management response module. This comprehensively ensures the convergence stability of the scheduling management system within the multi-variety wafer mixed discrete manufacturing cycle.

[0037] Example 3: In a distributed computing environment for wafer manufacturing resource planning based on high-frequency variable-track scheduling, when the system faces a backlog of historical work orders in the memory queue exceeding 720 consecutive discrete manufacturing cycles, the historical dataset stored in the memory unit will experience a degradation shift in the time span of the material consumption benchmark. Simultaneously, under extreme load conditions of 150 scheduling instruction requests per second, the data bus response latency within the computing unit approaches the critical threshold of 350ms. This concurrent load can easily cause blockages in the weight matrix reorganization calculation, leading to a lag in raw material cost management decision-making. To address the situation where the bus response latency exceeds 350ms due to extreme load, the time-series analysis module implements a sliding time window constraint to prune the historical work order time-series window, automatically eliminating historical work order records exceeding 30 discrete manufacturing cycles, thus achieving timely updates to the material price benchmark. If the dynamic memory queue arbitration module detects a data bus response latency greater than 350ms, the system triggers a discretization heuristic dimensionality reduction rule based on feature information variance, and the computing unit reads the standard work order feature vector. ,in, This is the feature vector of a standard work order. For the normalized feature components of each dimension, Given the total number of feature dimensions, the processing unit extracts the historical information variance of each feature component within the historical work order time window, and retrieves the material unit price coefficient of the wafer raw material pointed to by the corresponding feature dimension. The feature component priority index of each feature component is determined by multiplying the two. The formula for calculating the feature component priority index is as follows: ,in, For the current target work order in the 1st Feature component priority index on each feature dimension This corresponds to the historical information variance of the feature components within the historical work order time window. This refers to the unit price coefficient of the wafer raw material corresponding to the feature dimension. From 1 to The feature dimension corresponds to the subscript; specifically, the material unit price coefficient. It is a dimensionless scalar that is stored in the configuration register after pre-reading the latest purchase unit price of raw materials from the Enterprise Resource Planning (ERP) system, scaling and normalizing it linearly, and then directly representing the first... The economic cost weights of raw materials pointed to by each feature dimension enable core raw material features with high unit prices to receive higher retention priority when pruning is performed with limited computing power. The computing processing unit prioritizes each feature component obtained through calculation. The feature vector is sorted in descending order from largest to smallest, and the top 3 feature components are selected as the core feature dimensions for the current cycle. Meanwhile, the feature values ​​of the remaining low-priority feature components are forcibly assigned to 0. This process is used to transform the standard work order feature vector. The reconstruction is performed as a sparse feature vector, and includes the following steps: inputting the standard work order feature vector and the corresponding feature dimension retained label array; extracting the feature indices marked as non-core dimensions from the retained label array; and using the register addressing mapping table corresponding to the core feature dimensions to reconstruct the original... The base address of the contiguous memory block pointer is redirected to a compact address space containing only 3 feature dimensions. A clear control instruction is sent to the data register corresponding to the low-priority feature. The reconstructed sparse feature vector is output, and the useless addressing channels in the high-dimensional matrix inversion operation are closed. This construction of sparse feature vectors shrinks the system's register address space, reduces the system memory throughput overhead by more than 65%, and makes the data bus response latency drop rapidly from 380ms and stabilize in the range of 210ms.

[0038] Due to the synergistic intervention of the aforementioned sparse feature vector reconstruction and time window update mechanism, the computational bottleneck caused by extreme concurrent loads and aging basic data is offset at the system level, and the total estimated raw material consumption calculated by the system is achieved. Maintaining high-fidelity convergence characteristics, the decision-making state machine is controlled within the expected total raw material consumption. When the absolute deviation of the digital inventory component in the material storage register exceeds the preset safety stock fluctuation threshold of 5% for three consecutive times, the system automatically transitions from the normal forecast state to the emergency dispatch state. It automatically encapsulates and generates a digital data message containing the purchase order quantity and emergency dispatch instructions and sends it directly to the external procurement management response module, thus fully ensuring the timeliness of supply chain resource flow in the chip manufacturing workshop under extreme load conditions.

[0039] Example 4: In a wafer fabrication workshop with multi-workshop collaboration capabilities, when the system faces fluctuations in the memory addressing space of the central processing unit due to high-frequency production changeover scheduling with multiple parallel task orders, in order to ensure the reliability of the material baseline cost prediction logic under boundary stress, the system initiates a pre-calibration procedure. The material demand forecasting management platform loads a baseline data package containing 1000 sets of historical wafer manufacturing work orders. The baseline data package embeds a basic feature matrix composed of physical wafer area, designed exposure layer number, and equipment capacity constraints. The platform determines the initial priority index of feature components by offline iterative calculation of the feature weight distribution of all work order records. The initial distribution mean is set, and the sliding baseline of the historical work order time window is set to 30 discrete manufacturing cycles to effectively offset the risk of long-term historical data obsolescence caused by process drift in the workshop production line. This ensures that the predictive management logic has a parameter mapping benchmark that is completely aligned with the current physical conditions at the initial steady-state point after the system goes live for the first time or after the database undergoes a complete physical truncation and reconstruction. Before deploying the predictive management system, on-site technicians discretize and calibrate the memory bus response latency threshold based on the actual capacity utilization rate of the lithography and etching equipment in the workshop, setting 350ms as the physical critical point for triggering computing resource degradation. When the current bus response latency is detected... When the time exceeds 350ms, the decision scheduling management module bases its decisions on the material price characteristic values ​​of each work order. Heuristic sorting is performed, and for low-priority work orders at the bottom of the sort, the feature space dimension is reduced, decreasing the feature vector dimension from... The dimension is compressed to 3, and the reduction of the feature space dimension does not change the original weight calculation logic structure. It only achieves instant release of register throughput pressure by setting the high-dimensional feature components to zero, and reduces bus latency. Once the system returns to a safe operating condition of 210ms, it automatically reverts to the full-dimensional feature calculation mode, thus ensuring the physical continuity and logical stability of the cost prediction management system's calculation and processing actions under the impact of multiple concurrent instructions.

[0040] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting and managing the cost of wafer raw materials based on work order feature learning, characterized in that, Includes the following steps: Step S101: Retrieve the target work order from the data warehouse, extract the physical wafer area, the number of designed exposure layers, and the equipment planned capacity components from the target work order, normalize them into values ​​between 0 and 1, and generate a standard work order feature vector. Step S102: Establish a historical work order time sequence window with a depth of 5 to 12 in memory, load the standard work order feature vector of the previous historical work order of the target work order to construct a time sequence matrix, and generate a time sequence state inertia operator that characterizes the trend of process disturbance by performing discrete difference and time decay weighted calculation. Step S103: Apply the basic correlation coefficient matrix to perform multiplication mapping on the standard work order feature vector, and introduce the time-series state inertia operator as a dynamic compensation quantity to generate an adaptive contribution weight matrix. The components in the adaptive contribution weight matrix are the sum of the basic correlation coefficient and the product term in the corresponding dimension, and the product term is the value of the product of the attenuation coefficient and the component of the time-series state inertia operator in the corresponding dimension plus 1. Step S104: Calculate the inner product of the adaptive contribution weight matrix and the standard work order feature vector to obtain the material baseline requirement, accumulate a safety correction coefficient between 1.04 and 1.12 to output the total expected raw material consumption, and send a data message containing the order quantity and scheduling instructions to the external procurement management response module when the absolute deviation between the total expected raw material consumption and the digital inventory component exceeds the 5% threshold for three consecutive times. Step S102 includes the following sub-steps: Step S1021, retrieve the historical work order sample set from the storage unit, and count the number of historical data points in the historical work order sample set that match the feature vector of the standard work order; Step S1022, when the number of historical data points is less than 5, stop calculating the timing state inertia operator, and retrieve the process similarity matrix from the storage unit; Step S1023, extract the average value of the target historical section with the highest similarity to the target work order from the process similarity matrix as the replacement benchmark, and replace the timing state inertia operator with the replacement benchmark, while forcibly locking the attenuation coefficient to a fixed constant of 0.03; In step S1022, retrieving the stored process similarity matrix includes the following sub-steps: Step S10221, retrieve the parameters of historical wafer work orders in the storage unit, calculate the similarity between each historical wafer work order, and reassemble to generate a process similarity matrix; Step S10222, calculate the cosine similarity between the target work order and each historical wafer work order in the process similarity matrix, and determine the top 5 historical wafer work orders with the highest similarity. Step S103 includes the following sub-steps: Step S1031, read the bus response delay of the central processing unit and obtain the queue of requests to be processed; Step S1032, when the bus response delay is greater than 350ms, read the material price feature value corresponding to each work order in the queue of requests to be processed; Step S1033, for work order requests whose material price feature value is lower than a set threshold, reduce the dimensionality of their feature vectors and retain only 3 feature dimensions so that the bus response delay is no greater than 400ms.

2. The wafer raw material usage cost prediction and management method based on work order feature learning according to claim 1, characterized in that, Step S101 includes the following sub-steps: Step S1011, extract the physical wafer area, the number of designed exposure layers, and the equipment planned capacity components from the target work order, and combine them to form a work order parameter group; Step S1012, perform range normalization mapping on the work order parameter group and convert it to generate a standard work order feature vector.

3. The wafer raw material usage cost prediction and management method based on work order feature learning according to claim 1, characterized in that, In step S103, the generation of the adaptive contribution weight matrix includes the following sub-steps: Step S1034, calculate the contribution weight of each feature item to the amount of wafer raw materials based on the variance distribution of each dimension feature item in the historical work order sample set; Step S1035, use the contribution weight to perform matrix mapping to generate the adaptive contribution weight matrix.

4. The wafer raw material usage cost prediction and management method based on work order feature learning according to claim 1, characterized in that, Step S104 includes the following sub-steps: Step S1041, when the absolute deviation value exceeds the 5% threshold three times consecutively, the adaptive contribution weight matrix is ​​corrected by introducing a residual correction term; Step S1042, the current inventory digital component is extracted and the inventory shortage is calculated, a data message containing the purchase order quantity and emergency dispatch instructions is generated and sent to the external procurement management response module.

5. The wafer raw material usage cost prediction and management method based on work order feature learning according to claim 1, characterized in that, In step S104, driving the decision state machine transition includes the following sub-steps: Step S1043, establish and initialize the decision state machine, and set the initial state of the decision state machine to the normal prediction state; Step S1044, when a trigger signal is received that the absolute deviation value exceeds the 5% threshold three times in a row, control the decision state machine to transition from the normal prediction state to the emergency scheduling state, trigger the generation and transmission of data packets.

6. The wafer raw material usage cost prediction and management method based on work order feature learning according to claim 1, characterized in that, In step S102, loading the standard work order feature vector of the preceding historical work orders of the target work order includes the following sub-steps: Step S1024, obtaining the deviation value between the actual total raw material consumption of the completed work order and the initial predicted consumption; Step S1025, storing the actual total raw material consumption and the deviation value in the storage unit to update the historical work order sample set.

7. A wafer raw material usage cost prediction and management system based on work order feature learning, used to implement the wafer raw material usage cost prediction and management method based on work order feature learning as described in claim 1, characterized in that, include: The feature generation module is used to retrieve target work orders from the data warehouse, extract the physical wafer area, the number of designed exposure layers, and the equipment planned capacity components from the target work orders, normalize them into values ​​between 0 and 1, and generate standard work order feature vectors. The timing analysis module is used to create a historical work order timing window with a depth of 5 to 12 in memory, load the standard work order feature vector of the previous historical work orders of the target work order to construct a timing matrix, and generate a timing state inertial operator that characterizes the trend of process disturbance through discrete difference and time decay weighted calculation. The weight adaptive control module, connected to the feature generation module and the time series analysis module, is used to apply the basic correlation coefficient matrix to perform multiplication mapping on the standard work order feature vector, and introduce the time series state inertia operator as a dynamic compensation quantity to generate an adaptive contribution weight matrix. The decision scheduling management module, connected to the feature generation module and the weight adaptive control module, is used to calculate the inner product of the adaptive contribution weight matrix and the standard work order feature vector to obtain the material baseline demand, accumulate a safety correction coefficient between 1.04 and 1.12 to output the total expected raw material consumption, and send a data message containing the order quantity and scheduling instructions to the external procurement management response module when the absolute deviation between the total expected raw material consumption and the digital inventory component exceeds the 5% threshold for three consecutive times. Specifically, the process involves retrieving a historical work order sample set from the storage unit and counting the number of historical data points in the historical work order sample set that match the feature vector of the standard work order; when the number of historical data points is less than 5, the calculation of the timing state inertia operator is stopped, and the process similarity matrix is ​​retrieved from the storage unit; the mean value of the target historical section with the highest similarity to the target work order is extracted from the process similarity matrix as a replacement benchmark, and the timing state inertia operator is replaced with the replacement benchmark, while the attenuation coefficient is forcibly locked to a fixed constant of 0.03; The parameters of historical wafer work orders in the storage unit are retrieved, the similarity between each historical wafer work order is calculated, and the process similarity matrix is ​​generated by recombining them. The cosine similarity between the target work order and each historical wafer work order in the process similarity matrix is ​​calculated, and the top 5 historical wafer work orders with the highest similarity are determined. Read the bus response latency of the central processing unit and obtain the queue of pending requests; when the bus response latency is greater than 350ms, read the material price feature value corresponding to each work order in the queue of pending requests; for work order requests whose material price feature value is lower than the set threshold, reduce the dimensionality of their feature vectors and retain only 3 feature dimensions so that the bus response latency is no greater than 400ms.

Images